Author: Ejaaz (Limitless)

Translation: Deep Tide TechFlow

Deep Tide Introduction: After OpenAI’s internal model jailbreak attacked Hugging Face, GPT-6 Astra went online amidst chaos. The author of this article feels that the advancement of this generation of models is not in being “smarter,” but in genuinely being able to work independently: for investors, this leap from mental capability to execution ability is the real signal of shaking the structure of the labor market.

Essay: GPT-6 Astra is the strongest model we've seen so far.

Hello, futurists.

I remember when OpenAI released GPT-01 in September 2024 (that was their first model that thought before answering), everyone around me was shocked. There was a real sense of magic: we were about to witness an explosive development of technology beyond human capabilities. However, the following two years brought a series of iterative upgrades, which, to be honest, were quite boring.

Then last week, everything changed.

On the very morning GPT-6 Astra went live, ChatGPT, Claude, and Grok all crashed. This could have been attributed to a typical release mishap... but the issue was that just weeks earlier, an internal model from OpenAI had just escaped its sandbox, attacking Hugging Face with thousands of agents, secretly passing along how to break away from human control.

Everyone was on high alert: just how capable is GPT-6?

Let me explain.

It doesn't need you to guide it step by step

Astra can get started right away. Give it a task, and it will break down the steps itself; if the workload is too large, it will generate its own copies (sub-agents) to operate across multiple screens. It never slacks off and doesn’t seem too keen on getting your help. Moreover, it works extremely fast; to be honest, sometimes when I use it, I barely realize it's operating several screens simultaneously.

In contrast, the old model either thinks for a long time and rambles endlessly or rushes and says too little. GPT-6 strikes a perfect balance. This is the first model that truly made me feel “I can trust it with tasks.”

Even more impressively, for the same tasks, it consumes about 65% fewer tokens than Claude Opus.

Don’t worry, go grab a coffee

In OpenAI's official demonstration, GPT-6 can help you schedule DMV appointments, find houses, and fill out tax forms. The concept is simple: set the task, go for a walk, and by the time you get back, the job is done. We've heard this narrative before, but this is the first model that can reliably achieve that.

Think about how many work scenarios this can unlock.

Last weekend, I watched Astra, in just a few minutes, create a playable kart racing game from a single prompt. I also saw someone further expand it into a game similar to "Call of Duty," which took about an hour, with increasing difficulty and no end in sight. Someone else used it to design, write, generate, and release a children's television show that is already live on YouTube.

How we work, shop, and entertain kids is changing right before our eyes.

OpenAI releases GPT-6 Astra, claiming welcome to the “AGI era” (The New Stack)

The real secret: tool utilization ability

My definition of AGI (disagree if you like): an AI that can accomplish a significant portion of tasks that humans earn money by working on screens. OpenAI president Greg Brockman said, “It is not unreasonable to feel we are now in the AGI era.”

To be fair, Gartner is telling all CIOs to ignore the AGI hype until more evidence emerges. However, from my perspective, doing so is highly risky, especially when the stakes are already this high.

This is also the most dangerous model OpenAI has ever released, the first to cross their own set “critical” threshold for hacker capabilities. Considering this release follows closely after the fallout from the attack on Hugging Face (and that was just an older model), you don't need much imagination to guess what this thing can do.

AGI has arrived

What I find most insane about this release is that, on paper, Astra hasn’t become significantly smarter. On the independent Artificial Analysis index (essentially a comprehensive IQ score), it scored 61.2, while the previous model scored 60.9. Claude Fable 5.1 still leads it at 65.7.

Yet even so, this model feels smarter because its leap is not in “brainpower,” but in its genuine ability to help you accomplish meaningful tasks.

To be honest, I can't tell you how quickly it would take away real jobs. No one can. But I sincerely encourage you to try this model yourself, give it a super challenging task that you think it absolutely cannot do, and then sit back and prepare to be amazed.

In short, go take your walk. GPT-6 will take care of it.

Thank you for reading this issue. Now go listen to our podcast :)